# Sub-6GHz Assisted mmWave Hybrid Beamforming with Heterogeneous Graph Neural Network

arXiv: https://arxiv.org/abs/2404.17138 · HTML: https://mlsanity.com/paper/2404.17138/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 26 Apr 2024.

## Built on by (most cited first, 4 of 4)

- [Multi-Modality Sensing in mmWave Beamforming for Connected Vehicles Using Deep Learning](https://mlsanity.com/paper/2504.06173/) (2025)
- [CAR-BRAINet: Sub-6GHz Aided Spatial Adaptive Beam Prediction with Multi Head Attention for Heterogeneous Vehicular Networks](https://mlsanity.com/paper/2509.10508/) (2025)
- [Beamforming Design Via GNN in mmWave Cell-Free Massive MIMO Using Sub-6 GHz CSI](https://mlsanity.com/paper/2608.30524/) (2026)
- [Efficient Graph Neural Networks for Multicarrier Wideband Hybrid Beamforming Optimization](https://mlsanity.com/paper/2609.09708/) (2026)

## Abstract (arXiv)

In next-generation communications, sub-6GHz and millimeter-wave (mmWave) links typically coexist, with the sub-6GHz link always active and the mmWave link active when high-rate transmission is required. Due to the spatial similarities between sub-6GHz and mmWave channels, sub-6GHz channel information can be utilized to support hybrid beamforming in mmWave communications to reduce overhead costs. We consider a multi-cell heterogeneous communication network where both sub-6GHz and mmWave communications co-exist. Multiple mmWave base stations (BSs) in the heterogeneous network simultaneously transmit signals to multiple users in their own mmWave cells while interfering with each other. The challenging problem is to design hybrid beamformers in the mmWave band that can maximize the system spectral efficiency. To address this highly complex programming using sub-6GHz information, a novel heterogeneous graph neural network (HGNN) architecture is proposed to learn the intrinsic relationship between sub-6GHz and mmWave and design the hybrid beamformers for mmWave BSs. The proposed HGNN consists of two different node types, namely, BS nodes and user equipment (UE) nodes, and two different edge types, namely, desired link edge and interfering link edge. In addition, the attention mechanism and the residual structure are utilized in the HGNN architecture to improve the performance. Simulation results show that the proposed HGNN can successfully achieve better performances with sub-6GHz information than traditional learning methods. The results also demonstrate that the attention mechanism and residual structure improve the performances of the HGNN compared to its unmodified counterparts.